Abstract
Graph neural networks (GNNs), as topology/structure-aware models within deep learning, have emerged as powerful tools for AI-aided drug discovery (AIDD). By directly operating on molecular graphs, GNNs offer an intuitive and expressive framework for learning the complex topological and geometric features of drug-like molecules, cementing their role in modern molecular modeling. This review provides a comprehensive overview of the methodological foundations and representative applications of GNNs in drug discovery, spanning tasks such as molecular property prediction, virtual screening, molecular generation, biomedical knowledge graph construction, and synthesis planning. Particular attention is given to recent methodological advances, including geometric GNNs, interpretable models, uncertainty quantification, scalable graph architectures, and graph generative frameworks. We also discuss how these models integrate with modern deep learning approaches, such as self-supervised learning, multi-task learning, meta-learning and pre-training. Throughout this review, we highlight the practical challenges and methodological bottlenecks encountered when applying GNNs to real-world drug discovery pipelines, and conclude with a discussion on future directions.
Introduction
Bringing a new drug to market typically costs over $2 billion and takes more than a decade. Historically, "Computer-Aided Drug Design" relied on human-engineered molecular fingerprints or simplified string representations like SMILES. However, these methods often fail to capture the 3D spatial orientation or the complex topological relationships that dictate how a drug interacts with a human protein. Graph Neural Networks (GNNs) address this by treating molecules as mathematical graphs where atoms are nodes and bonds are edges.
Methodological Foundations
The landscape of GNN methodologies for drug discovery is built upon several key pillars:
- Spatial vs. Spectral GNNs: Early models utilized spectral theory (graph signal processing), but the field has largely shifted toward Message Passing Neural Networks (MPNNs). These spatial models update an atom's state by aggregating information from its neighbors, effectively mimicking local chemical environments.
- Symmetric and Geometric GNNs: To respect physical laws, models must be invariant (predicting the same energy regardless of rotation) or equivariant (predicting forces that rotate with the molecule). This evolution has moved from simple distance-based models like SchNet to higher-order angular models such as DimeNet and GemNet.
- Graph Transformers: Leveraging global attention mechanisms, Graph Transformers allow every atom to interact with every other atom in a molecule. This architecture helps mitigate the "over-squashing" problem where information is lost in large, complex molecular structures.
- Generative Paradigms: There is a notable shift from Variational Autoencoders (VAEs) to Diffusion Models and Flow Matching. These are currently state-of-the-art for generating 3D molecular conformations and designing ligands that fit precisely into protein binding pockets.
Data Resources and Benchmarking
The effectiveness of these models relies on robust datasets. Key benchmarks include:
- Property Prediction: QM9 (quantum mechanics), ZINC (drug-like molecules), and Tox21 (toxicity).
- Virtual Screening: PDBBind and BioLiP for protein-ligand interactions.
- Synthesis Planning: USPTO and Reaxys.
- Emerging Benchmarks: MUBen for uncertainty quantification and CBGBench for structure-based molecular generation.
Key Insights and Performance
Modern GNNs have reached several significant milestones. Graph Transformers (e.g., Specformer, Graphormer) have pushed Mean Absolute Error (MAE) on the ZINC dataset down to the 0.06–0.12 range. Meanwhile, geometric models like GemNet have achieved a 2x performance improvement in predicting atomic forces, which is critical for molecular dynamics simulations.
Beyond raw performance, the review highlights that 2D graphs are no longer sufficient; successful models must incorporate 3D coordinates and physical symmetries (SO(3) and E(3) groups). Furthermore, GNNs are exhibiting "scaling laws" similar to Large Language Models, where pretraining on billions of unlabeled molecules significantly boosts performance on smaller, high-quality experimental datasets. Finally, explainability remains a non-negotiable requirement in drug design, as chemists need substructure attribution to understand the specific parts of a molecule responsible for predicted effects.
Challenges and Future Outlook
Despite rapid progress, several bottlenecks remain. Data quality is a primary concern, as experimental noise and batch effects can hinder model reliability. Generative models still occasionally produce physically implausible molecules with unrealistic bond lengths or steric clashes. Additionally, "negative transfer" can occur when pretraining on general datasets hurts performance on specific downstream tasks.
The future of the field lies in System-Level Modeling. Moving beyond single-molecule analysis, researchers are developing Biomedical Knowledge Graphs that connect molecules to genes, diseases, and clinical phenotypes, aiming to understand drug mechanisms within the context of the entire human body.
Summary updated